A naukri.com initiative
ML News
Arxiv
1d
146
Image Credit: Arxiv
Temporal and Semantic Evaluation Metrics for Foundation Models in Post-Hoc Analysis of Robotic Sub-tasks
- Recent works in Task and Motion Planning (TAMP) show that training control policies on language-supervised robot trajectories with quality labeled data improves task success rates.
- A framework is presented to decompose trajectory data into temporally bounded and natural language-based sub-tasks.
- An algorithm named SIMILARITY is introduced to measure the temporal alignment and semantic fidelity of language descriptions in sub-task decompositions.
- The framework demonstrates high scores for both temporal similarity and semantic similarity, above 90%, compared to a randomized baseline of 30% in multiple robotic environments.
Read Full Article
8 Likes
Arxiv
1d
346
Image Credit: Arxiv
A predictive machine learning force field framework for liquid electrolyte development
- A predictive machine learning force field framework for liquid electrolyte development.
- Introducing BAMBOO (ByteDance AI Molecular Simulation Booster), a predictive framework for molecular dynamics (MD) simulations for liquid electrolyte in lithium batteries.
- Utilizes a physics-inspired graph equivariant transformer architecture to learn from quantum mechanical simulations.
- Demonstrates state-of-the-art accuracy in predicting key electrolyte properties such as density, viscosity, and ionic conductivity.
Read Full Article
20 Likes
Arxiv
1d
256
Image Credit: Arxiv
FastLloyd: Federated, Accurate, Secure, and Tunable $k$-Means Clustering with Differential Privacy
- We study the problem of privacy-preserving $k$-means clustering in the horizontally federated setting.
- Existing federated approaches using secure computation suffer from substantial overheads and do not offer output privacy.
- The work provides enhancements to both differentially private (DP) and secure computation components to achieve better speed, privacy, and accuracy.
- By utilizing the computational DP model, a lightweight, secure aggregation-based approach is designed, achieving significant speed improvement and improved utility.
Read Full Article
15 Likes
Arxiv
1d
331
Image Credit: Arxiv
GP-MoLFormer: A Foundation Model For Molecular Generation
- GP-MoLFormer is an autoregressive molecular string generator trained on over 1.1 billion chemical SMILES.
- It performs well on three different generative tasks: de novo generation, scaffold-constrained molecular decoration, and property-guided optimization.
- GP-MoLFormer demonstrates its general utility and compares favorably to existing baselines.
- The model shows strong memorization of training data, impacted by the quality and scale of the training data.
Read Full Article
19 Likes
Arxiv
1d
304
Image Credit: Arxiv
Hierarchical Procedural Framework for Low-latency Robot-Assisted Hand-Object Interaction
- Advances in robotics have led to the development of human-robot interaction (HRI) technologies.
- A hierarchical procedural framework is proposed to enable dynamic robot-assisted hand-object interaction.
- The framework leverages computer vision (CV) for 3D reconstruction of the human hand and motion primitives for robotic actions.
- Experimental validation demonstrates the effectiveness of the hierarchical control architecture, achieving a delay of ≤ 0.3 seconds in tele-interaction scenarios.
Read Full Article
18 Likes
Arxiv
1d
150
Image Credit: Arxiv
Eliminating Position Bias of Language Models: A Mechanistic Approach
- Position bias is a prevalent issue in modern language models.
- The bias leads to unexpected model failures and affects performance, robustness, and reliability.
- A mechanistic analysis identifies causal attention and relative positional encodings as the sources of bias.
- A training-free zero-shot approach called PINE (Position-INvariant inferencE) is proposed to eliminate the bias and improve performance in downstream tasks.
Read Full Article
9 Likes
Arxiv
1d
33
Image Credit: Arxiv
One Policy to Run Them All: an End-to-end Learning Approach to Multi-Embodiment Locomotion
- Deep Reinforcement Learning techniques are achieving state-of-the-art results in robust legged locomotion.
- URMA (Unified Robot Morphology Architecture) is introduced as a framework to control different embodiments of legged robots.
- The framework utilizes an end-to-end Multi-Task Reinforcement Learning approach and morphology-agnostic encoders and decoders.
- Experiments show that URMA can learn a locomotion policy that can be transferred to unseen robot platforms in simulation and the real world.
Read Full Article
2 Likes
Arxiv
1d
128
Image Credit: Arxiv
Privacy Vulnerabilities in Marginals-based Synthetic Data
- Privacy vulnerabilities have been identified in marginals-based synthetic data generation.
- Marginals-based synthetic data generation algorithms leak information about individuals that can be recovered more efficiently than previously understood.
- A membership inference attack, MAMA-MIA, has been developed to exploit these vulnerabilities.
- The attack allows for more accurate and faster learning of hidden data compared to other leading attacks.
Read Full Article
7 Likes
Arxiv
1d
33
Image Credit: Arxiv
Assessing and improving reliability of neighbor embedding methods: a map-continuity perspective
- Neighbor embedding methods, such as t-SNE and UMAP, are widely used for visualizing high-dimensional data.
- A lack of data-independent notions of embedding maps in these methods can introduce misleading visual artifacts.
- Researchers have introduced LOO-map, a framework that extends embedding maps to the entire input space, aiming to improve reliability.
- Two types of diagnostic scores have been developed to detect unreliable embedding points and improve hyperparameter selection.
Read Full Article
2 Likes
Arxiv
1d
79
Image Credit: Arxiv
1-2-3-Go! Policy Synthesis for Parameterized Markov Decision Processes via Decision-Tree Learning and Generalization
- A learning-based approach is proposed to synthesize policies for huge parameterized Markov decision processes (MDPs).
- The method generalizes optimal policies obtained from model-checking small instances to larger ones using decision-tree learning.
- By bypassing the need for explicit state-space exploration of large models, the method provides a practical solution to the state-space explosion problem.
- Experimental results show that the policies perform well even for models beyond the reach of state-of-the-art analysis tools.
Read Full Article
4 Likes
Arxiv
1d
131
Image Credit: Arxiv
Patient-specific prediction of glioblastoma growth via reduced order modeling and neural networks
- Researchers present a mathematical model for predicting the growth of glioblastoma (GBL) and identifying patient-specific parameters from neuroimaging data.
- The model utilizes a diffuse-interface mathematical model and a reduced-order modeling strategy trained on synthetic data derived from patient-specific brain anatomies reconstructed from imaging.
- A neural network surrogate learns the inverse mapping from tumor evolution to model parameters, resulting in significant computational speed-up while maintaining high accuracy.
- The study establishes a foundation for the development of patient-specific digital twins in neuro-oncology for future clinical applications.
Read Full Article
7 Likes
Arxiv
1d
94
Image Credit: Arxiv
Multi-objective Combinatorial Methodology for Nuclear Reactor Site Assessment: A Case Study for the United States
- As clean energy demand grows, repurposing coal power plant sites (CPP) with existing infrastructure is one way to reduce costs for nuclear power plants (NPP).
- A multi-objective optimization methodology, using combinatorial search, evaluated over 30,000 potential NPP sites in the United States.
- The methodology generated a comprehensive database of site locations, attributes, site scores, and the contribution of each attribute to the score.
- Results indicate that CPP sites in Ohio, North Carolina, and New Hampshire, as well as Brownfield sites in Florida and California, are promising locations for nuclear development.
Read Full Article
5 Likes
Arxiv
1d
169
Image Credit: Arxiv
Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales
- Researchers propose a method for fully quantized Winograd convolution to reduce computational and storage costs in large-scale text-to-image diffusion models.
- Quantization of diffusion models has been explored in previous works to reduce compute costs and memory bandwidth usage.
- The proposed method focuses on finer-grained group-wise quantization, combined with finetuning the scale parameters of the Winograd transform matrices.
- The method achieves near-lossless quality in text-to-image generation and outperforms state-of-the-art methods in image classification tasks.
Read Full Article
10 Likes
Arxiv
1d
131
ZETA: Leveraging Z-order Curves for Efficient Top-k Attention
- The Transformer model, widely used for sequence modeling, relies on self-attention which becomes inefficient for long sequences.
- The ZETA method proposes leveraging Z-Order Curves for Efficient Top-k Attention to address the inefficiency of self-attention.
- ZETA enables parallel querying of past tokens and achieves similar performance to self-attention while reducing computational demands.
- Experimental results show that ZETA outperforms attention models on various language modeling tasks.
Read Full Article
7 Likes
Medium
1d
299
Top 5 Books to Learn AI: In My Point of View
- 1. 'Artificial Intelligence: A Modern Approach' by Russell and Norvig is the go-to guide for understanding AI fundamentals.
- 2. 'Deep Learning' by Goodfellow, Bengio, and Courville explains the mechanics of neural networks and their applications.
- 3. 'The Master Algorithm' by Pedro Domingos explores the concept of a universal algorithm for deriving knowledge from data.
- 4. 'Superintelligence' by Nick Bostrom delves into the risks and rewards of advanced AI and its alignment with human goals.
Read Full Article
17 Likes
For uninterrupted reading, download the app